Coal mill fault early warning method fusing quantum weighting GRU and multi-mode classification

Abstract

The invention relates to a coal mill fault early warning method fusing quantum weighting GRU and multi-modal classification, and the method comprises the steps: firstly, proposing a data fusion architecture based on multi-equipment cooperative perception, integrating coal mill body parameters, associated equipment operation indexes and environment monitoring data, constructing a dynamic coupling feature library, and strengthening the representation capability of complex working conditions of a production area; secondly, a quantum weighted GRU neural network is designed, a quantum information processing module is embedded in a gating mechanism, the extraction precision of nonlinear time sequence features is remarkably improved, and signal noise in a high-dust environment is suppressed in combination with an adaptive filtering technology; an environmental parameter dynamic correction mechanism is further introduced, a residual threshold is adaptively adjusted according to real-time temperature and humidity and dust concentration, and the adaptability of the model to extreme working conditions is enhanced; and finally, accurate discrimination of multiple fault types such as coal interruption, coal blockage, spontaneous combustion and the like is realized through a hybrid architecture of a multi-classification model, and high-robustness and high-precision fault early warning support is provided for a thermal power generating unit production area.

Description

TECHNICAL FIELD

[0001] The present application relates to coal mill fault early warning, in particular to a coal mill fault early warning method fusing quantum weighted GRU and multi-modal classification. BACKGROUND

[0002] The current coal mill and coal feeder fault early warning technology in the production area of thermal power generating units generally has problems such as high data isolation, sensitive environmental interference and insufficient model generalization ability. The traditional method mainly relies on single sensor data (such as current or temperature) of the coal mill itself, and does not effectively integrate the cooperative operation parameters of the associated equipment in the production area (such as the primary air fan and the sealing fan), resulting in insufficient representation of fault characteristics. The existing algorithm model (such as standard GRU or SVM) has limited processing capability for high-noise and nonlinear time series data in the production area, is difficult to capture the dynamic coupling relationship of multiple devices, and lacks an adaptive adjustment mechanism for complex working conditions such as high temperature and high dust, resulting in a high false alarm rate. In addition, most technologies do not construct a multi-source data fusion framework, and cannot correct the model threshold in combination with environmental parameters (such as temperature, humidity, and dust concentration), which restricts the robustness and practicality of the early warning system. Therefore, improvement and innovation are imperative. SUMMARY

[0003] In view of the above situation, in order to overcome the defects of the prior art, the purpose of the present application is to provide a coal mill fault early warning method fusing quantum weighted GRU and multi-modal classification, which can effectively solve the problem of high-precision coal mill fault early warning.

[0004] The technical solution solved by the present application is:

[0005] A coal mill fault early warning method fusing quantum weighted GRU and multi-modal classification, comprising the following steps:

[0006] Step 1: Data acquisition and preprocessing

[0007] 1) DCS real-time data: Collect coal mill operating parameters, including current, inlet and outlet differential pressure, outlet temperature, vibration value, motor power, with a sampling frequency of 1 second;

[0008] 2) Associated equipment: Read the sealing air pressure, primary air volume, and coal feeder speed operating data, with a sampling frequency of 1 second;

[0009] 3) Current environmental information (low-frequency data): Obtain environmental temperature, humidity, air pressure, and carbon monoxide concentration;

[0010] Step 2: Time series alignment and interpolation

[0011] The low-frequency data of environmental temperature, humidity, and air pressure are collected once a minute, and are aligned to 1 second time series using cubic spline interpolation:

[0012]

[0013] wherein:

[0014] x aligned (t): the value of the aligned time series data at time point t;

[0015] a i : the i-th coefficient of the cubic spline interpolation polynomial;

[0016] t: time variable, unit: second;

[0017] t k ,t k+1 : time endpoints of the interpolation interval;

[0018] Step three: data cleaning and denoising

[0019] The vibration signal of the coal mill vibration value is filtered in the frequency domain, the effective frequency band is retained, and high-frequency noise interference is eliminated:

[0020]

[0021] wherein:

[0022] X filtered (f): the amplitude of the filtered vibration signal at frequency f;

[0023] X(f): the frequency domain amplitude of the original vibration signal;

[0024] Step four: statistical correlation screening

[0025] According to the Pearson correlation coefficient, the correlation coefficient r of each parameter and the fault label is calculated, and the strongly correlated parameters with |r|>0.3 are retained:

[0026]

[0027] wherein:

[0028] r is the Pearson correlation coefficient, with a value range of [-1, 1], and the greater the absolute value, the stronger the correlation;

[0029] x i is the parameter value of the i-th sample;

[0030] y i is the fault label of the i-th sample, 0-normal, 1-fault;

[0031] is the mean value of the parameter and the fault label;

[0032] Seal air pressure, coal feeder speed are excluded, and five core parameters of current, primary air volume, inlet and outlet differential pressure, vibration value, and outlet temperature are retained;

[0033] Step five: quantum-weighted GRU model construction

[0034] (1) Input feature design:

[0035] The input vector X(t) = [I(t), Q(t), ΔP(t), A(t), T(t)] is constructed, with a dimension of 5 x 60

[0036] In the formula:

[0037] I(t) is the current of the coal mill;

[0038] Q(t) is the primary air volume;

[0039] ΔP(t) is the inlet and outlet differential pressure;

[0040] A(t) is the vibration value;

[0041] T(t) is the outlet temperature of the coal mill;

[0042] (2) Quantum gating mechanism implementation:

[0043] Embedding quantum-weighted neurons in the update gate (z t ) and reset gate (r t ) of GRU:

[0044]

[0045] In the formula:

[0046] z t is the output of the GRU update gate at time step t;

[0047] σ is the Sigmoid activation function;

[0048] W z is the weight matrix of the update gate;

[0049] h t-1 is the hidden state of the previous time step;

[0050] x t is the input vector of the current time step;

[0051] b z is the bias term of the update gate;

[0052] α, β are the coefficients of the quantum superposition state, satisfying α 2 + β 2 = 1;

[0053] |0>, β|1> are the ground states of the quantum bits (basic states in quantum computing);

[0054] is a tensor product operator, representing the fusion of quantum states and traditional signals.

[0055] The quantum gradient descent algorithm is used to update the weights, and the loss function is the mean square error (MSE) of the predicted value and the actual value;

[0056] (3) Anti-interference module design:

[0057] Add a noise estimation network after the input layer, and output the corrected features through residual learning:

[0058]

[0059] In the formula:

[0060] X clean is the denoised feature vector;

[0061] X raw is the original input feature vector;

[0062] is a noise estimation network;

[0063] Step six: dynamic parameter optimization and model training

[0064] (1) The quantum weighted GRU model contains multiple hyperparameters. If unbounded search is used, the consumption of computing resources will increase exponentially;

[0065] Therefore, the hyperparameter search space is defined:

[0066] Hidden layer node number: N h ∈[50, 200], to avoid too few or too many nodes;

[0067] Initial learning rate: η ∈ [0.001, 0.1], to prevent the gradient descent step size from being too large or too small;

[0068] Quantum gate initial parameter α0 ∈ [0.6, 0.9], to ensure the effectiveness of the quantum superposition state;

[0069] By defining the search space, the optimization calculation amount is reduced by 70%, and the SSA-PSO hybrid algorithm can converge within 30 minutes;

[0070] (2) SSA-PSO hybrid optimization algorithm

[0071] Particle swarm optimization (PSO):

[0072] Particle position update formula:

[0073] v i+1 = wv i +c1r1(p best-x i )+c2r2(g best -x i )

[0074] wherein:

[0075] v i+1 is the update speed of the next generation of particles;

[0076] w is the inertia weight, which controls the influence of the historical speed;

[0077] c1, c2 are acceleration constants, which represent the weights of individual and group experience, respectively;

[0078] r1, r2: random numbers in the interval [0, 1];

[0079] p best is the individual historical optimal position of the particle;

[0080] g best is the group historical optimal position;

[0081] x i is the current position of the particle.

[0082] Sparrow Search Algorithm (SSA):

[0083] Sparrow position update rule:

[0084]

[0085] Termination condition: verification set MSE changes <0.1% for 10 consecutive generations;

[0086] wherein:

[0087] is the position of the k+1th generation of sparrows;

[0088] x best is the current optimal position of the population;

[0089] β is the random step size coefficient of the follower;

[0090] α is the decay coefficient of the discoverer;

[0091] K is the maximum number of iterations;

[0092] (3) Model training

[0093] Use historical 3-month normal data to train quantum weighted GRU, batch size 128, number of iterations 500;

[0094] The convergence time of the optimized model is ≤30 minutes;

[0095] Step seven: multi-modal residual analysis and fault classification

[0096] (1) Residual definition and classification

[0097] Residual is the difference between actual observation and model prediction, which is used to quantify prediction error or equipment operation anomaly, and residual is divided into two categories:

[0098] 1) Time series prediction residual: prediction error based on quantum weighted GRU model;

[0099] 2) Environmental compensation residual: correct the influence of environmental factors on threshold, used for dynamic threshold adjustment;

[0100] (2) Time series residual calculation

[0101] Real-time calculation of prediction residual: e(t) = x 实际 (t) - x 预测 (t)

[0102] In the formula:

[0103] x 实际 (t) is the parameter collected from the sensor in real time;

[0104] x 预测 (t) is the output of the quantum weighted GRU model;

[0105] e(t) is the time series residual value;

[0106] (3) Environmental compensation residual calculation

[0107] Temperature compensation residual: w 温度 = k T (T 实际 -T 基准 )

[0108] In the formula:

[0109] k T is the temperature compensation coefficient, specifically 0.03 / ℃;

[0110] T 实际 is the actual collected temperature;

[0111] T 基准 is the reference temperature, specifically 20℃;

[0112] Humidity compensation residual: e 湿度 = k H (H 实际 H 基准 )

[0113] In the formula:

[0114] k HTc is temperature compensation coefficient, specifically 0.01 / %RH;

[0115] H 实际 Tc is temperature compensation coefficient, specifically 0.01 / %RH;

[0116] H 基准 Tc is temperature compensation coefficient, specifically 0.01 / %RH;

[0117] (4) Residual fusion and dynamic threshold

[0118] Combine three types of residuals into a vector: E = [e 时序 , e 温度 , e湿度 ]

[0119] Calculate the time series residual mean μ and standard deviation α based on the sliding window (60 seconds), and the dynamic threshold is set as:

[0120] T(t) = μ + 3α + e 温度 + e 湿度

[0121] In the formula:

[0122] μ is the mean of time series residuals in the sliding window;

[0123] α is the standard deviation of time series residuals in the sliding window;

[0124] e 温度 is temperature compensation;

[0125] e 温度 is humidity compensation;

[0126] T(t) is the dynamic threshold.

[0127] (5) Residual decision rule

[0128]

[0129] (2) Multi-classification model implementation

[0130] 1) Train special classifiers for different fault types, output preliminary probability, SVM sub-model training: coal cutting classifier: input features are current deviation, primary air volume deviation;

[0131] Coal blocking classifier: input features are inlet and outlet pressure difference, vibration value;

[0132] Self-ignition classifier: input is outlet temperature, CO concentration;

[0133] 2) Use the classifier probability output model to convert the SVM decision value into probability value of each fault;

[0134] (3) LightGBM integrated classification:

[0135] 1) Input feature construction

[0136] SVM probability vector: P = [P 断煤 , P 堵煤 , P 自燃 ]

[0137] Residual vector: E = [e 时序 , e 温度 , e 湿度 ]

[0138] Combined input:

[0139] Where:

[0140] P 断煤 is the preliminary coal breaking probability output by the SVM model after training;

[0141] P 堵煤 is the preliminary coal blocking probability output by the SVM model after training;

[0142] P 自燃 is the preliminary spontaneous combustion probability output by the SVM model after training;

[0143] e 时序 is the time series prediction residual output by the step seven GRU model;

[0144] e 温度 is the temperature compensation residual output by the step seven GRU model;

[0145] e 温度 is the humidity compensation residual output by the step seven GRU model;

[0146] 2) Decision rule

[0147] Through the LightGBM decision classification model, output four types of probability:

[0148]

[0149] The final label is:

[0150]

[0151] Where:

[0152] P 断煤 is the final coal breaking probability output by the LightGBM decision classification model after training;

[0153] P 堵煤 is the final coal blocking probability output by the LightGBM decision classification model after training;

[0154] Final spontaneous combustion probability output after training of LightGBM decision classification model;

[0155] y is the final failure type;

[0156] Step seven: graded early warning triggering and output

[0157] (1) Confidence evaluation:

[0158] Generate probability distribution based on historical normal data kernel density estimation, set 95% confidence interval:

[0159]

[0160] In the formula:

[0161] F(x) is the cumulative distribution function;

[0162] KDE(u) is the kernel density estimation function, which generates probability density based on historical normal data;

[0163] F -1 (0.95) is the threshold value corresponding to the cumulative probability of 95%;

[0164] Only when the failure probability P i > T 95％ Trigger alarm;

[0165] (2) Graded alarm logic

[0166] Take parameter deviation and confidence interval as the basis for graded alarm, where parameter deviation is the relative deviation percentage of actual value and predicted value, and confidence is the failure probability output by the mixed classification model;

[0167] Early warning level: single parameter deviation > 15% & confidence > 80%, push to DCS operation interface, remind the operator to continue monitoring;

[0168] Severe level: multi-parameter deviation > 25% & confidence > 95%, trigger audible and visual alarm, immediately stop the coal mill;

[0169] Emergency level: residual error exceeds threshold value for 3 times in a row & spontaneous combustion probability > 99%, trigger CO2 fire extinguishing system;

[0170] (3) Online model update

[0171] Add false positives / false negatives cases to the training set and trigger incremental learning:

[0172]

[0173] In the formula:

[0174] θnew θ old are the updated and pre-updated model parameters;

[0175] η is a learning rate, which controls the parameter update step size;

[0176] is the gradient of the loss function with respect to the parameters;

[0177] X new , y new are the new training data and its label.

[0178] The present application overcomes the limitations of the existing technology in multiple dimensions: first, a data fusion architecture based on multi-device collaborative perception is proposed, which integrates the parameters of the coal mill, the operating indicators of associated devices and environmental monitoring data, builds a dynamic coupling feature library, and enhances the representation ability of complex working conditions in the production area; second, a quantum weighted GRU neural network is designed, which embeds a quantum information processing module in the gating mechanism, significantly improves the extraction accuracy of nonlinear time series features, and combines adaptive filtering technology to suppress signal noise in high dust environments; further introduce the environmental parameter dynamic correction mechanism, according to the real-time temperature and humidity and dust concentration to adaptively adjust the residual threshold, enhance the adaptability of the model to extreme working conditions; finally, through the mixed architecture of multi-classification model, the precise discrimination of coal cutting, coal blocking, spontaneous combustion and other fault types is realized, providing high robustness and high precision fault warning support for the production area of thermal power generating units. The present scheme solves the core defects of traditional methods in data collaboration, environmental adaptability and classification ability, and provides reliable technical support for the intelligent operation and maintenance of thermal power generation. BRIEF DESCRIPTION OF DRAWINGS

[0179] Figure 1 is a flowchart of the method of the present application. DETAILED DESCRIPTION

[0180] The specific implementation of the present application is further described in detail below in conjunction with the embodiments.

[0181] As shown in Figure 1 , the present application is a coal mill fault warning method that fuses quantum weighted GRU and multi-modal classification, including the following steps:

[0182] Step 1: Data acquisition and preprocessing

[0183] 1) DCS real-time data: collect coal mill operating parameters, including current, inlet and outlet differential pressure, outlet temperature, vibration value, motor power, sampling frequency 1 second;

[0184] 2) Associated equipment: read the sealing air pressure, primary air volume, coal feeder speed operating data, sampling frequency 1 second;

[0185] 3) Current environmental information (low-frequency data): Obtain ambient temperature, humidity, air pressure, and carbon monoxide concentration to correct the heat dissipation efficiency model of the coal mill;

[0186] Step two: Time alignment and interpolation

[0187] The low-frequency data of ambient temperature, humidity, and air pressure are collected every minute, and cubic spline interpolation is used to align them to 1 second time sequence:

[0188]

[0189] In the formula:

[0190] x aligned (t): The value of the aligned time sequence data at time point t;

[0191] a i : The i-th term coefficient of the cubic spline interpolation polynomial;

[0192] t: Time variable, unit is second (s);

[0193] t k ,t k+1 : Time endpoints of the interpolation interval;

[0194] The timestamp synchronization error is controlled within 10 ms to ensure the time consistency of multi-source data;

[0195] Step three: Data cleaning and denoising

[0196] The vibration signal of the coal mill vibration value is filtered in the frequency domain, the effective frequency band (200-600 Hz) is retained, and the high-frequency noise interference is eliminated:

[0197]

[0198] In the formula:

[0199] X filtered (f): The amplitude of the filtered vibration signal at frequency f;

[0200] X(f): The frequency domain amplitude of the original vibration signal;

[0201] Step four: Statistical correlation screening

[0202] According to the Pearson correlation coefficient, the correlation coefficient r of each parameter and the fault label is calculated, and the strongly correlated parameters with |r|>0.3 are retained:

[0203]

[0204] In the formula:

[0205] r is the Pearson correlation coefficient, the value range is [-1, 1], the greater the absolute value, the stronger the correlation;

[0206] x i is the parameter value (such as current value) of the i-th sample;

[0207] y i is the fault label of the i-th sample, 0-normal, 1-fault;

[0208] is the mean value of the parameter and the fault label;

[0209] Sealing air pressure, coal feeder speed are excluded, and current, primary air volume, inlet and outlet differential pressure, vibration value, outlet temperature are retained as five core parameters;

[0210] Step five: quantum weighted GRU model construction

[0211] (1) Input feature design:

[0212] The input vector (t) = [I(t), Q(t), ΔP(t), A(t), T(t)] is constructed, and the dimension is 5x60 (60-second sliding window)

[0213] In the formula:

[0214] I(t) is the mill current;

[0215] Q(t) is the primary air volume;

[0216] ΔP(t) is the inlet and outlet differential pressure;

[0217] A(t) is the vibration value;

[0218] T(t) is the mill outlet temperature;

[0219] (2) Quantum gate mechanism implementation:

[0220] Embed quantum weighted neurons in the update gate (z t ) and reset gate (r t ) of GRU:

[0221]

[0222] In the formula:

[0223] z t is the output of the GRU update gate at time step t;

[0224] σ is the Sigmoid activation function;

[0225] W z is the weight matrix of the update gate;

[0226] h t-1 is the hidden state of the previous time step;

[0227] x t is the input vector of the current time step;

[0228] b z is the bias term of the update gate;

[0229] α, β are coefficients of the quantum superposition state, satisfying α 2 + β 2 = 1;

[0230] |0>, β|1> are the ground states of the quantum bits (basic states in quantum computing);

[0231] is the tensor product operator, representing the fusion of the quantum state and the traditional signal.

[0232] The quantum gradient descent (QGD) algorithm is used to update the weights, and the loss function is the mean square error (MSE) between the predicted value and the actual value;

[0233] (3) Anti-interference module design:

[0234] A noise estimation network is added after the input layer, and the corrected features are output through residual learning:

[0235]

[0236] where:

[0237] X clean is the denoised feature vector;

[0238] X raw is the original input feature vector;

[0239] is the noise estimation network (fully connected neural network);

[0240] Step six: dynamic parameter optimization and model training

[0241] (1) The quantum weighted GRU model contains multiple hyperparameters (such as hidden layer node number, learning rate, quantum gate initial parameters), and if unbounded search is used, the computational resource consumption will grow exponentially;

[0242] Therefore, the hyperparameter search space is defined:

[0243] Hidden layer node number: N h ∈ [50, 200], to avoid too few nodes (underfitting) or too many nodes (overfitting);

[0244] Initial learning rate: η ∈ [0.001, 0.1], to prevent the gradient descent step too large (shock) or too small (slow convergence);

[0245] Quantum gate initial parameter a0∈[0.6, 0.9], to ensure the effectiveness of quantum superposition state;

[0246] By defining the search space, the optimization calculation is reduced by 70%, and the SSA-PSO hybrid algorithm can converge within 30 minutes;

[0247] (2) SSA-PSO hybrid optimization algorithm

[0248] Particle swarm optimization (PSO):

[0249] Particle position update formula:

[0250] v i+1 = wv i + c1r1(p best -x i ) + c2r2(g best -x i )

[0251] In the formula:

[0252] v i+1 is the updated speed of the next generation of particles;

[0253] w is the inertia weight, which controls the influence of historical speed;

[0254] c1, c2 are acceleration constants, representing the weights of individual and group experience respectively;

[0255] r1, r2: random numbers in the interval [0, 1];

[0256] p best is the individual historical optimal position of the particle;

[0257] g best is the group historical optimal position;

[0258] x i is the current position of the particle.

[0259] Sparrow search algorithm (SSA):

[0260] Sparrow position update rule:

[0261]

[0262] Termination condition: verification set MSE changes less than 0.1% for 10 generations in a row;

[0263] In the formula:

[0264] is the position of the sparrow in the k+1k+1 generation;

[0265] x best is the optimal position of the current population;

[0266] β is the random step coefficient of the follower;

[0267] α is the attenuation coefficient of the finder;

[0268] K is the maximum number of iterations;

[0269] (3) Model training

[0270] Use the historical 3-month normal data to train the quantum weighted GRU with a batch size of 128 and 500 iterations;

[0271] The optimized model convergence time is ≤30 minutes (traditional GRU requires ≥50 minutes);

[0272] Step 7: Multimodal residual analysis and fault classification

[0273] (1) Definition and classification of residuals

[0274] Residuals are the difference between actual observations and model predictions (or theoretical values), and are used to quantify prediction errors or equipment operation anomalies. In this technical solution, residuals are divided into two categories:

[0275] 1) Time series prediction residual: prediction error based on the quantum weighted GRU model;

[0276] 2) Environmental compensation residual: corrects the impact of environmental factors on the threshold and is used for dynamic threshold adjustment;

[0277] (2) Calculation of time series residuals

[0278] Real-time calculation of prediction residual: e(t) = x 实际 (t)-x 预测 (t)

[0279] Where:

[0280] x 实际 (t) is the parameter collected from the sensor in real time (e.g., the actual current value is 120A);

[0281] x 预测 (t) is the output of the quantum weighted GRU model (e.g., the predicted current value is 118A);

[0282] e(t) is the time series residual value (for example, if the actual current value is 120A and the predicted current value is 118A, the time series residual value is 2A);

[0283] (3) Temperature compensation residual error calculation

[0284] Temperature compensation residual error: e 温度 = k T (T 实际 - T 基准 )

[0285] Wherein:

[0286] k T is the temperature compensation coefficient, specifically 0.03 / ℃;

[0287] T 实际 is the actual collection temperature;

[0288] T 基准 is the reference temperature, specifically 20℃;

[0289] Humidity compensation residual error: e 湿度 = k H (H 实际 - H 基准 )

[0290] Wherein:

[0291] k H is the temperature compensation coefficient, specifically 0.01 / %RH;

[0292] H 实际 is the actual collection temperature;

[0293] H 基准 is the reference temperature, specifically 40%RH;

[0294] (4) Residual error fusion and dynamic threshold

[0295] The three types of residual errors are combined into a vector: E = [e 时序 , e 温度 , e 湿度 ] The time series residual mean μ and standard deviation α are calculated based on a sliding window (60 seconds), and the dynamic threshold is set as:

[0296] T(t) = μ + 3α + e 温度 + e 湿度

[0297] Wherein:

[0298] μ is the mean of the time series residual within the sliding window;

[0299] α is the standard deviation of the time series residual within the sliding window;

[0300] e 温度 is the temperature compensation;

[0301] e 温度For humidity compensation;

[0302] T(t) is a dynamic threshold.

[0303] (5) Residual decision rule

[0304]

[0305] (2) Multi-classification model implementation

[0306] 1) Train a special classifier for different fault types, output preliminary probability, SVM sub-model training:

[0307] Coal breaking classifier: input features are current deviation, primary air volume deviation, kernel function is hybrid kernel (RBF + linear); coal blocking classifier: input features are inlet and outlet pressure difference, vibration value, kernel function is polynomial kernel (order 3); spontaneous combustion classifier: input is outlet temperature, CO concentration (indirectly calculated), kernel function is RBF kernel function;

[0308] 2) Convert the SVM decision value into probability value of each fault using the classifier probability output model;

[0309] (3) LightGBM integrated classification:

[0310] 1) Input feature construction

[0311] SVM probability vector: P = [P 断煤 , P 堵煤 , P 自燃]

[0312] Residual vector: E = [e 时序 , e 温度 , e 湿度 ]

[0313] Combined input: (Dimension is 3+3 vector)

[0314] In the formula:

[0315] P 断煤 is the preliminary coal breaking probability output after SVM model training;

[0316] P 堵煤 is the preliminary coal blocking probability output after SVM model training;

[0317] P 自燃 is the preliminary spontaneous combustion probability output after SVM model training;

[0318] e 时序 is the time series prediction residual output by step seven GRU model;

[0319] e 温度Temperature-compensated residual output by the Step 7 GRU model

[0320] e 温度 Humidity-compensated residual output by the Step 7 GRU model

[0321] 2) Decision rule

[0322] Output four types of probabilities through the LightGBM decision classification model:

[0323]

[0324] The final label is:

[0325]

[0326] Where:

[0327] P 断煤 Final coal-out probability output by the LightGBM decision classification model after training;

[0328] P 堵煤 Final coal-out probability output by the LightGBM decision classification model after training;

[0329] Final spontaneous combustion probability output by the LightGBM decision classification model after training;

[0330] y is the final fault type.

[0331] Step 7: Hierarchical early warning triggering and output

[0332] (1) Confidence evaluation:

[0333] Generate probability distribution based on historical normal data kernel density estimation (KDE), and set 95% confidence interval:

[0334]

[0335] Where:

[0336] F(x) is the cumulative distribution function;

[0337] KDE(u) is the kernel density estimation function, which generates probability density based on historical normal data;

[0338] F -1 (0.95) is the threshold value corresponding to the cumulative probability of 95%;

[0339] Only when the fault probability P i > T 95％ trigger an alarm;

[0340] (2) Hierarchical alarm logic

[0341] The parameter deviation degree and the confidence interval are used as the basis of the hierarchical alarm, wherein the parameter deviation degree is the relative deviation percentage of the actual value and the predicted value (or the rated value), and the confidence is the fault probability output by the hybrid classification model;

[0342] Early warning level: single parameter deviation degree > 15% & confidence > 80%, pushed to the DCS operation interface, reminding the operator to continue monitoring;

[0343] Severe level: multi-parameter deviation degree > 25% & confidence > 95%, triggering sound and light alarm, and the operator immediately stops the coal mill;

[0344] Emergency level: residual error continuously exceeds the threshold value for 3 times & self-ignition probability > 99%, linkage CO2 fire extinguishing system;

[0345] (3) Online model updating

[0346] Adding false positives / false negatives cases to the training set, triggering incremental learning:

[0347]

[0348] In the formula:

[0349] θ new , θ old are the model parameters after updating and before updating;

[0350] η is the learning rate, controlling the parameter update step;

[0351] is the gradient of the loss function to the parameter;

[0352] X new , y new are the new training data and its label.

[0353] The method of the present application is compared with the existing traditional method, and the results are as follows:

[0354] The traditional coal blocking alarm algorithm relies on the DCS differential pressure alarm (the threshold value is fixed at 200 kPa), and when the actual coal blocking occurs, the differential pressure rises to 220 kPa to trigger the alarm, but at this time the coal powder has been seriously accumulated, and it takes 4 hours to clean up.

[0355] As shown in the following table, the present application realizes a breakthrough improvement in the fault early warning of the coal mill and coal feeder of the thermal power unit by fusing the quantum weighted GRU model, dynamic feature selection and multi-source collaborative perception mechanism: it innovatively uses quantum gate to enhance the time series modeling capability, combines with automatic feature screening to remove redundant noise, so that the fault detection accuracy is significantly improved, and the residual error is reduced by 33% compared with the traditional method; through the mixed classification model and the environment adaptive threshold design, it accurately distinguishes multiple types of faults (such as coal breaking, coal blocking, spontaneous combustion), and the false positive rate is reduced to below 2% under complex working conditions such as high temperature and high dust; at the same time, relying on edge computing and dynamic optimization algorithm, the early warning response time is less than 2 seconds, and online self-learning of the model is supported to adapt to equipment aging, finally reducing the operation and maintenance cost by more than 30% while ensuring real-time, providing a high-robustness, high-precision intelligent operation and maintenance solution for thermal power generation systems.

[0356]

Claims

1. A coal mill fault warning method integrating quantum weighted GRU and multimodal classification, characterized in that: The following steps are involved: Step 1: Data collection and preprocessing 1) DCS real-time data: collects coal mill operating parameters, including current, inlet and outlet differential pressure, outlet temperature, vibration value, and motor power, with a sampling frequency of 1 second; 2) Related equipment: read the operating data of sealing air pressure, primary air volume, and coal feeder speed, with a sampling frequency of 1 second; 3) Current environmental information: obtain ambient temperature, humidity, air pressure, and carbon monoxide concentration; Step 2: Timing alignment and interpolation Low-frequency data such as ambient temperature, humidity, and air pressure are collected once a minute and aligned to a 1-second time series using cubic spline interpolation: Where: x aligned (t): the value of the aligned time series data at time point t; a i : the i-th coefficient of the cubic spline interpolation polynomial; t: time variable, in seconds; t k ,t k+1 : The time endpoints of the interpolation interval; Step 3: Data cleaning and denoising Perform frequency domain filtering on the vibration signal of the coal mill vibration value to retain the effective frequency band and eliminate high-frequency noise interference: Where: X filtered (f): amplitude of the filtered vibration signal at frequency f; X(f): frequency domain amplitude of the original vibration signal; Step 4: Statistical correlation screening According to the Pearson correlation coefficient, the correlation coefficient r between each parameter and the fault label is calculated, and the strongly correlated parameters with |r|>0.3 are retained: Where: r is the Pearson correlation coefficient, which ranges from [-1,1]. The larger the absolute value, the stronger the correlation. x i is the parameter value of the i-th sample; y i is the fault label of the i-th sample, 0-normal, 1-fault; is the mean of the parameter and the fault label; Eliminate sealing air pressure and coal feeder speed, and retain the five core parameters: current, primary air volume, inlet and outlet differential pressure, vibration value, and outlet temperature; Step 5: Quantum Weighted GRU Model Construction (1) Input feature design: Construct input vector X(t) = [I(t), Q(t), ΔP(t), A(t), T(t)], dimension = 5×60 Where: I(t) is the coal mill current; Q(t) is the primary air volume; ΔP(t) is the differential pressure between inlet and outlet; A(t) is the vibration value; T(t) is the coal mill outlet temperature; (2) Implementation of quantum gating mechanism: In the GRU update gate (z t ) and reset gate (r t ) is embedded with quantum weighted neurons: Where: z t Update the output of the GRU gate at time step t; σ is the Sigmoid activation function; W z is the weight matrix of the update gate; h t-1 is the hidden state of the previous time step; x t is the input vector of the current time step; b z is the bias term of the update gate; α, β are the coefficients of quantum superposition state, satisfying α 2 +β 2 =1; |0>, β|1> is the ground state of the quantum bit (the basic state in quantum computing); is a tensor product operator, representing the fusion of quantum states and traditional signals. The quantum gradient descent algorithm is used to update the weights, and the loss function is the mean square error (MSE) between the predicted value and the actual value; (3) Anti-interference module design: Add a noise estimation network after the input layer and output the corrected features through residual learning: Where: X clean is the eigenvector after denoising; X raw is the original input feature vector; is the noise estimation network; Step 6: Dynamic parameter optimization and model training (1) The quantum weighted GRU model contains multiple types of hyperparameters. If unbounded search is used, the computing resource consumption will increase exponentially; Therefore, the hyperparameter search space is defined as: Number of hidden layer nodes: N h ∈[50,200], avoid too few or too many nodes; Initial learning rate: η∈[0.001,0.1], to prevent the gradient descent step from being too large or too small; The initial parameter of the quantum gate α0∈[0.6,0.9] ensures the validity of the quantum superposition state; By defining the search space, the optimization computational effort was reduced by 70%, and the SSA-PSO hybrid algorithm converged within 30 minutes. (2) SSA-PSO hybrid optimization algorithm Particle Swarm Optimization (PSO): Particle position update formula: v i+1 =wv i +c1r1(p best -x i )+c2r2(g best -x i ) Where: v i+1 The update speed of the next generation of particles; w is the inertia weight, which controls the influence of historical velocity; c1 and c2 are acceleration constants, representing the weights of individual and group experience, respectively; r1, r2: random numbers in the interval [0,1]; p best is the optimal historical position of the individual particle; g best is the historical optimal position of the group; x i is the current position of the particle. Sparrow Search Algorithm (SSA): Sparrow position update rules: Termination condition: the MSE of the validation set changes less than 0.1% for 10 consecutive generations; Where: is the position of the sparrow in the k+1k+1 generation; x best is the optimal position of the current population; β is the random step coefficient of the follower; α is the attenuation coefficient of the finder; K is the maximum number of iterations; (3) Model training Use the historical 3-month normal data to train the quantum weighted GRU with a batch size of 128 and 500 iterations; The convergence time of the optimized model is ≤30 minutes; Step 7: Multimodal residual analysis and fault classification (1) Definition and classification of residuals Residuals are the difference between the actual observed value and the model predicted value. They are used to quantify the prediction error or equipment operation anomaly. Residuals are divided into two categories: 1) Time series prediction residual: prediction error based on the quantum weighted GRU model; 2) Environmental compensation residual: corrects the impact of environmental factors on the threshold and is used for dynamic threshold adjustment; (2) Calculation of time series residuals Real-time calculation of prediction residual: e(t) = x 实际 (t)-x 预测 (t) Where: x 实际 (t) is the parameter collected from the sensor in real time; x 预测 (t) is the output of the quantum weighted GRU model; e(t) is the time series residual value; (3) Calculation of environmental compensation residuals Temperature compensation residual: e 温度 =k T (T 实际 -T 基准 ) Where: k T is the temperature compensation coefficient, specifically 0.03 / °C; T 实际 is the actual collected temperature; T 基准 is the reference temperature, specifically 20°C; Humidity compensation residual: e 湿度 =k H (H 实际 -H 基准 ) Where: k H is the temperature compensation coefficient, specifically 0.01 / %RH; H 实际 is the actual collected temperature; H 基准 is the reference temperature, specifically 40% RH; (4) Residual fusion and dynamic threshold Combine the three types of residuals into a vector: E = [e 时序 , e 温度 , e 湿度 ] The mean μ and standard deviation α of the time series residual are calculated based on the sliding window (60 seconds), and the dynamic threshold is set as: T(t)=μ+3α+e 温度 +e 湿度 Where: μ is the mean of the time series residuals in the sliding window; α is the standard deviation of the time series residual within the sliding window; e 温度 For temperature compensation; e 湿度 To compensate for humidity; T(t) is the dynamic threshold. (5) Residual determination rules (2) Implementation of multi-classification model 1) Train dedicated classifiers for different fault types and output preliminary probabilities. SVM sub-model training: Coal failure classifier: Input features are current deviation and primary air volume deviation; Coal blocking classifier: Input features are inlet and outlet pressure difference and vibration value; Spontaneous combustion classifier: input is outlet temperature and CO concentration; 2) Use the classifier probability output model to convert the SVM decision value into the probability value of each fault; (3) LightGBM ensemble classification: 1) Input feature construction SVM probability vector: P = [P 断煤 , P 堵煤 , P 自燃 ] Residual vector: E = [e 时序 , e 温度 , e 湿度 ] Combination input: X LightGBM =P⊕E Where: P 断煤 It is the initial coal failure probability output after SVM model training; P 堵煤 It is the preliminary coal blocking probability output after SVM model training; P 自燃 It is the preliminary spontaneous combustion probability output after SVM model training; e 时序 The time series prediction residual output by the GRU model in step 7; e 温度 The temperature compensation residual output by the GRU model in step 7; e 湿度 The humidity compensation residual output by the GRU model in step 7; 2) Decision-making rules Through the LightGBM decision classification model, four types of probabilities are output: The final label is: Where: P 断煤 The final coal failure probability output by the LightGBM decision classification model after training; P 堵煤 The final coal blocking probability output by the LightGBM decision classification model after training; The final spontaneous combustion probability output after LightGBM decision classification model training; y is the final fault type. Step 7: Tiered warning triggering and output (1) Confidence assessment: Generate a probability distribution based on the kernel density estimate of historical normal data and set a 95% confidence interval: Where: F(x) is the cumulative distribution function; KDE(u) is the kernel density estimation function, which generates probability density based on historical normal data; F -1 (0.95) is the threshold corresponding to the cumulative probability of 95%; Only when the failure probability P i >T 95％ When the alarm is triggered; (2) Hierarchical alarm logic The parameter deviation and confidence interval are used as the basis for graded alarms, where the parameter deviation is the relative deviation percentage between the actual value and the predicted value, and the confidence interval is the failure probability output by the hybrid classification model; Warning level: Single parameter deviation > 15% & confidence > 80%, the warning is pushed to the DCS operation interface to remind the operator to continue monitoring; Severe level: Multi-parameter deviation > 25% & confidence > 95%, triggering audible and visual alarms and immediately stopping the coal mill; Emergency level: If the residual exceeds the threshold value three times in a row and the probability of spontaneous combustion is greater than 99%, the CO2 fire extinguishing system will be activated; (3) Model online update Add false positive / missing cases to the training set to trigger incremental learning: Where: θ new ,θ old are the model parameters before and after the update; η is the learning rate, which controls the parameter update step size; is the gradient of the loss function with respect to the parameters; X new ,y new To add new training data and its labels.

Images